Multi-Dimensional Neural Mesh Compression via Joint 2D-3D Optimization

With the rapid development of 3D modeling technologies, high-precision 3D models are now widely used. However, the growth in model data size has brought serious challenges to storage, transmission, and rendering efficiency. To address these problems, a multi-dimensional optimized mesh compression method is proposed. In our work, a coarse mesh is first generated through geometric simplification. A feature field is then built on the coarse mesh, and spatial features at any point are obtained through bilinear interpolation. The position field and feature field are fed into a lightweight multilayer perceptron network, which predicts the displacement of each tessellated point to reconstruct fine details. A joint 2D–3D loss function is employed during optimization. The 2D rendering loss constrains image-level consistency through differentiable rendering, while the 3D normal loss corrects local geometry in three-dimensional space. In this way, both global shape and local details can be preserved even at high compression ratios. Experiments are conducted on the Stanford 3D Scanning Repository and Thingi10K datasets. The results show that the proposed method achieves a higher compression ratio and better geometric fidelity.